In the Netherlands, in the early years of this century, a small literary magazine maintained a digital dictionary called the Woordenboek. It was a reference for readers of contemporary Dutch literature - definitions of movements, terms of art, the vocabulary a literate person needed to navigate the conversation about books. The gatekeepers were human: critics with opinions, editors with taste, readers with letters to write. Whether a book was worth reading was a question answered by people who had read it.
Two decades later, the question has changed. The gatekeepers are algorithms. The question is no longer whether writing is good, but whether it is real. The dictionary has been replaced by a detection tool. The critic has been replaced by a probability score. The letter to the editor has been replaced by an automated flag.
The Old Gatekeepers
Literary criticism, for all its flaws, operated on a principle that now seems almost quaint: a human being reads a piece of writing and forms a judgment about it. That judgment might be wrong, biased, poorly argued, or brilliantly insightful. But it was always, fundamentally, an act of reading - a human mind encountering another human mind's work and responding to it.
The Dutch literary tradition that this site's predecessor participated in was particularly invested in this model. The Netherlands has one of Europe's highest rates of literary reading. Dutch literary criticism took itself seriously - reviewing a novel was not a consumer recommendation but an intellectual engagement. The Woordenboek existed because the conversation required shared vocabulary. You needed to know what a term meant before you could argue about whether a book embodied it.
This was gatekeeping, certainly. Not every book was reviewed. Not every writer was taken seriously. The system had all the usual problems of human judgment: favoritism, blind spots, institutional inertia, the tendency to mistake familiarity for quality. But the writers being evaluated were, at minimum, being read.
The New Gatekeepers
AI detection tools do not read. They calculate. When a tool like GPTZero or Turnitin AI processes a piece of text, it is performing statistical analysis on token sequences. It measures perplexity - how predictable the text is. It measures burstiness - how much variation exists in sentence complexity. It compares the text's statistical profile against models of what AI-generated text looks like.
The output is a number. A confidence score. A percentage. "This text is 87% likely to be AI-generated." There is no reading happening. No understanding. No engagement with what the text says, means, or accomplishes. The gatekeeper has been reduced to a classifier, and the classification is binary: human or machine.
This would be merely unfortunate if the stakes were low. But the stakes are careers, degrees, reputations, and livelihoods. When a university runs every student paper through a detector, when a publisher requires AI clearance before accepting a manuscript, when a client uses a free tool to check whether the freelancer they hired actually wrote the deliverable - the statistical classifier is making consequential decisions about human beings.
What We Lost in Translation
The literary critic could be wrong, but they could also be questioned. You could write a response. You could argue your case. You could point to the text itself and say, "Look at this passage - do you really think this is derivative?" The process, however imperfect, was dialogic. It involved human beings engaging with each other's reasoning.
The algorithm cannot be questioned in the same way. You cannot ask it to reconsider. You cannot point to a passage and explain what you meant. You cannot appeal to its experience as a reader, because it has none. When the number says 87%, the burden shifts entirely to you: prove the machine wrong. The conversation is over before it starts.
The Woordenboek was a tool for understanding. The AI detector is a tool for classification. The distance between those two functions is the distance between a culture that cares about writing and a culture that merely polices it.
A Third Way
There is nothing inevitable about the current model. We built it, and we can rebuild it. The technologies of content provenance - tools like C2PA that can create verifiable records of when, where, and how writing was produced - offer something neither the old critics nor the new algorithms provide: proof without judgment. Not "is this good?" or "is this real?" but "here is the documented history of how this was made."
Provenance does not replace literary criticism, and it does not replace critical reading. But it does replace the algorithmic gatekeeper with something more honest: a chain of evidence that the writer controls. It shifts the question from "what does the machine think?" to "what does the evidence show?" That shift - from statistical inference to documentary proof - is the most important change the writing world can make right now.
The Woordenboek helped readers understand books. Our glossary helps writers understand the tools being used to judge them. The continuity is not accidental. Understanding the vocabulary is still the first step toward participating in the conversation - and the conversation, in 2026, is about whether the human voice will continue to matter.